scGPT applicable to predicting effects of combinations of knockout mutations, instead of gain of function perturbations?

[swillus-august]

I intend to fine-tune scGPT in my graduate thesis using a single-cell RNA-seq CRISPR knockout screen (against tumor suppressors) to predict how the cells would respond if several genes were knocked out at once. In your 2024 scGPT Nature paper, you benchmarked this model using a gain-of-function CRISPR screen instead. Do you think that your verification approach for unseen perturbations (UMAP + Leiden) could also show clear clusters of cells that have the same knockouts?

Thank you very much for your time!

[shayanjl]

I think it is possible in principle, but I would be cautious about assuming that results from gain-of-function perturbations will automatically transfer to combinatorial knockouts. If your training data include enough knockout examples and the perturbation design is well represented, then UMAP and Leiden on predicted responses could still show clusters of cells with similar knockout combinations. But I would treat that as a hypothesis to test on knockout data, not something guaranteed by the original scGPT benchmark.